Abstract
Bird's-eye view (BEV) perception has gained significant attention because itprovides a unified representation to fuse multiple view images and enables awide range of down-stream autonomous driving tasks, such as forecasting andplanning. Recent state-of-the-art models utilize projection-based methods whichformulate BEV perception as query learning to bypass explicit depth estimation.While we observe promising advancements in this paradigm, they still fall shortof real-world applications because of the lack of uncertainty modeling andexpensive computational requirement. In this work, we introduce GaussianLSS, anovel uncertainty-aware BEV perception framework that revisitsunprojection-based methods, specifically the Lift-Splat-Shoot (LSS) paradigm,and enhances them with depth un-certainty modeling. GaussianLSS representsspatial dispersion by learning a soft depth mean and computing the variance ofthe depth distribution, which implicitly captures object extents. We thentransform the depth distribution into 3D Gaussians and rasterize them toconstruct uncertainty-aware BEV features. We evaluate GaussianLSS on thenuScenes dataset, achieving state-of-the-art performance compared tounprojection-based methods. In particular, it provides significant advantagesin speed, running 2.5x faster, and in memory efficiency, using 0.3x less memorycompared to projection-based methods, while achieving competitive performancewith only a 0.4% IoU difference.